PAIR is a router, not multi-GPU memory pooling

NVIDIA launched the beta of Personal AI Router at IFA 2026 as free open-source software for local AI inference.

Compatible machines discover each other on a local network and expose their inference engines through a common endpoint.

PAIR then sends independent requests to nodes that have capacity.

That distinction matters. Two graphics cards connected through PAIR do not suddenly share their VRAM or appear to software as one larger accelerator.

Every computer remains an independent node.

Multi-agent systems are the obvious target

Modern agent workflows often split one request into several tasks. One subagent may search documents, another can write code, while another checks the result.

Running all of them against one GPU creates a queue.

PAIR can dispatch those independent inference jobs to different computers instead.

In NVIDIA's five-subagent Hermes Desktop demonstration using Ollama, a workflow took approximately 18 minutes on one RTX Spark laptop.

The same workload finished in 8 minutes and 48 seconds when PAIR had three devices available.

That is a major reduction, but it is also a useful reminder that three machines do not automatically provide a perfect 3x speedup.

The router checks more than GPU utilization

PAIR's scheduler considers node readiness, inference-engine state, whether the requested model is already present and current GPU utilization.

NVIDIA uses mDNS for discovery and mTLS to protect communication between paired devices.

Prompts, files and agent context can remain inside the local network. Internet access is not required during operation once the necessary models have been downloaded.

Ollama and LM Studio are supported, allowing existing local-model workflows to use PAIR without rebuilding the entire agent harness.

An RTX 20-series card can participate

NVIDIA is not restricting the software to its newest hardware.

Supported nodes include GeForce RTX 20-series GPUs and newer, professional NVIDIA RTX GPUs based on Turing or later architectures, DGX Spark systems and Macs using Apple M4 silicon or newer.

Windows, Linux and macOS systems can therefore exist inside the same local inference pool.

That makes PAIR potentially more useful to people who already own several generations of hardware than to someone buying an entirely new cluster.

RTX Spark hardware arrives in October

NVIDIA also used IFA to put a more specific date on its new Windows platform. RTX Spark laptops and compact desktops are coming in October.

The top RTX Spark N1X configuration combines a 6,144-core Blackwell RTX GPU with a 20-core Grace CPU and up to 128GB of unified LPDDR5X memory.

NVIDIA rates it for up to 1 petaflop of FP4 AI compute.

A second configuration uses 5,120 GPU cores, an 18-core Grace CPU and up to 64GB of unified memory.

The listed power range is 45 to 80 watts depending on implementation.

128GB may matter more than the petaflop headline

Peak FP4 throughput produces the more impressive marketing number, but memory capacity is often the harder limitation for local AI.

A model that does not fit in available accelerator memory is not helped very much by another theoretical TOPS or FLOPS increase.

RTX Spark's unified architecture can expose as much as 128GB to CPU and GPU workloads rather than isolating AI inside a much smaller dedicated VRAM pool.

NVIDIA says the platform can run language models with as many as 120 billion parameters locally and support workflows with context windows reaching one million tokens under suitable conditions.

PAIR still does not add those memory pools together

PAIR can send ten inference jobs to several computers, but it does not automatically make a 128GB machine and a 64GB machine look like one 192GB accelerator.

A model that must be partitioned across several GPUs still needs a genuinely distributed inference system designed for that purpose.

PAIR is solving a different bottleneck: concurrent requests.

That happens to align extremely well with agent architectures, where a single user request can trigger many specialized workers at once.

NVIDIA is also claiming up to 1.9x faster llama.cpp inference

IFA brought backend improvements alongside the new router.

NVIDIA reports up to 1.9x higher llama.cpp throughput on a GeForce RTX 5090 through kernel optimizations, improved speculative decoding and faster prefill.

For vLLM, the company reports a 1.2x improvement on an RTX PRO 6000 Blackwell Workstation Edition and up to 1.4x on a two-DGX-Spark cluster.

Those are configuration-specific benchmark results, not a promise that every local model or every PC will suddenly become 1.9x faster.

A personal PC is starting to look like infrastructure

RTX Spark is NVIDIA's attempt to put workstation-class local AI capacity into ordinary Windows laptops and compact desktops.

PAIR extends that idea beyond one machine.

The interesting part is not really the phrase “home supercomputer.” Many enthusiasts and professionals already own multiple computers that spend much of the day underused.

NVIDIA is giving local agents a way to treat those idle machines as a shared queue of inference workers.

It will not replace cloud infrastructure or true distributed model serving. But for an agent trying to run five jobs at once, the old RTX PC in the next room may have just become useful again.